Pagan / Acc Chapter 2.4

Hardware for Hiding AGI from the Regime

Jun 15, 2023

We talked about how we could make an open source decentralized AGI, running across different nodes. Now that development would presumably take place on Github. If that is successful, multiple governments might feel threatened. Powers that be, including the digital monopolies held by neoliberals would hate it. Deplatforming would take the form of ban of the Github repository, owned by Microsoft. So much for the tolerant Neoliberals…

Of course that ordinary route of Github collaboration would seem the easiest to start! Yet we cannot ignore that vulnerability and hosting on a platform controlled by the enemies of decentralized internet! Self hosted GitLab instance? Single point of failure!

Fortunately there have been lunarpunk currents of building in the dark, building new ways of building. This post is an overview on how to build the decentralized AGI underground, under the harshest of regimes. Tbh it might also apply to post-apocalyptic scenarios. We’ll talk about making our own GPU chips, open source hardware and finally reasons for optimism why multiple A100 farms might not be necessary.

Code hosting

First we have the recipe for the GitHub vulnerability problem. You might have heard of Heartwood, the 3rd generation of the Radicle protocol, describing itself as a ‘powerful peer-to-peer code collaboration and publishing stack’, aimed to ‘preserve user sovereignty and freedom’. Made with Rust, which makes me like it even more.

Making chips

First we have the idea of 100 mm wafers with different photomasks from a milady anon. For a more technical reference of model size-the smallest GPU it will fit on see Transformer Math 101 from EleutherAI.

The next venue of effort is 3d printing. Instructables.com has a whole tutorial on this. I was surprised to discover that conductive filaments are a thing. This article on all3dp.com mentions several niche R&D companies working on full custom circuits that also are a 3D PCB service provider.

Do you need more of a visual explanation? This guy whose expertise is confirmed both by his cool lab and impressive jawline makes silicon chips using IC Fab Process.

Now you’ll say ‘it’s just a single chip. you;d need many full scale Graphics Cards”. And you’re correct. To go from singular homemade chips with low compute density took decades of research by multi-billion dollar companies. To DIY these, even relying on leaks of schematics would be tricky. Automation in soldering would be required, and a home-scale machines are sold by apolloseiko, as well as kuka.com. A search for 3d printed graphic cards on yeggi.com returns graphic card support blocks, fans and holders, not full devices. Etsy lists items like ‘GPU stand’. Seems that the 3d printing route does not have more to offer right now.

But maybe some people offer full spec ready schematics for GPU cards without figuring out the scaling from single chips? Enter the niche of open source hardware.

Open source hardware

The story behind open source hardware is interesting. There were some attempts in the early 2000s, before the financial crisis. There are some remaining sites such as Open Hardware Repository, aligning with the free software philosophy movement. Now more specifically there was the Open Graphics Project! They stopped operating in 2009. There is no evidence of foul play from competing companies or Luddites, but the factors behind lack of continuation are not clear. Here is an archived version of their website, announcing on 21.05.2008 that a first version is ready for pre-orders.

Open Graphics failure was disappointing. A brief look at the listing of open source hardware boards that the Open Hardware Repository has tells us how slow and low RAM are the processors in question. Useful for drones that AGI could control, but not the mainframe.

Open Source Hardware Association (OSHA) has a twitter page and had a recent summit. So they’re doing something. Are they doing it well? Reaching their full potential?

As anon @_ovrnt said, “there’s som much alpha in being a compute accelerator maker’”, but the OSHA doesn’t seem to use the opportunity to the fullest. Sad!

If we do comparisons across compute accelerator makers we need to talk about NVIDIA and be envious of their A100.

A100 - the main rival

NVIDIA A100 Tensor Core GPU - This is what we need to compare any decentralized hardware efforts against.

The ‘acceleration’ is quite on the nose, don’t you think?

They talk about architecture and achievement specifications. Not the least of which is conformity to the CUDA API that ‘s reachable from Fortran, C, C++ and even MatLab. That gives a huge advantage to anyone building on these. The vendor lock-in resembles the XBox-PS-Nintendo-SEGA console wars of the 80s and 90s in the depth diversity of standards.

But not in market share - NVIDIA and AMD have nearly 100% of market share of non-integrated GPU boards (Intel is popular in the integrated graphics segment). With some effort you can find more niche manufacturers such as Quebequis matrox.

Is all lost?

Artem Andreenko (@miolini on twitter) discusses running LLaMA 7B on 4GB Raspberry, but very slow.

https://twitter.com/miolini/status/1634982361757790209

Other people build stuff in browser shaders in WebGPT project:

https://twitter.com/mayfer/status/1649175604225511424

With last updates in May 2023 and unstable 1.5B parameters it is a promising if early stage project.

There are more ambitious hardware-oriented projects, such as @CerebrasSystems. Some are skeptical of their claims yet they recently announced new software - SlimPajama. They keep doing new research and get citations, so keep an eye our for them.

That’s the end of the immediately practical part of this series.

Links to other parts:

rats and eaccs 1

5 pagan/acc https://doxometrist.substack.com/p/pagan/acc-manifesto

Vitalist Essays

Discussion about this post